From b6353d5e221d11d281d6c9cdd0a049919622559d Mon Sep 17 00:00:00 2001
From: Yannick DAYER <yannick.dayer@idiap.ch>
Date: Wed, 12 May 2021 00:15:34 +0200
Subject: [PATCH] Set references in probes to prevent negative spoof
---
bob/bio/face/database/replaymobile.py | 105 ++++++++++++++------------
1 file changed, 57 insertions(+), 48 deletions(-)
diff --git a/bob/bio/face/database/replaymobile.py b/bob/bio/face/database/replaymobile.py
index 392609db..f48cac93 100644
--- a/bob/bio/face/database/replaymobile.py
+++ b/bob/bio/face/database/replaymobile.py
@@ -54,35 +54,6 @@ def load_frame_from_file_replaymobile(file_name, frame, should_flip):
image = numpy.transpose(image, (0, 2, 1))
return image
-def read_frame_annotation_file_replaymobile(file_name, frame, annotations_type="json"):
- """Returns the bounding-box for one frame of a video file of replay-mobile.
-
- Given an annnotation file location and a frame number, returns the bounding
- box coordinates corresponding to the frame.
-
- The replay-mobile annotation files are composed of 4 columns and N rows for
- N frames of the video:
-
- 120 230 40 40
- 125 230 40 40
- ...
- <x> <y> <w> <h>
-
- Parameters
- ----------
-
- file_name: str
- The annotation file name (relative to annotations_path).
-
- frame: int
- The video frame index.
- """
- logger.debug(f"Reading annotation file '{file_name}', frame {frame}.")
-
- video_annotations = read_annotation_file(file_name, annotation_type=annotations_type)
- # read_annotation_file returns an ordered dict with string keys
- return video_annotations[f"{frame}"]
-
class ReplayMobileCSVFrameSampleLoader(CSVToSampleLoaderBiometrics):
"""A loader transformer returning a specific frame of a video file.
@@ -101,39 +72,79 @@ class ReplayMobileCSVFrameSampleLoader(CSVToSampleLoaderBiometrics):
dataset_original_directory=dataset_original_directory,
)
self.reference_id_equal_subject_id = reference_id_equal_subject_id
+ self.references_list = []
def convert_row_to_sample(self, row, header):
"""Creates a set of samples given a row of the CSV protocol definition.
"""
- path = row[0]
- reference_id = row[1]
- id = row[2] # Will be used as 'key'
+ fields = dict([[str(h).lower(), r] for h, r in zip(header, row)])
- kwargs = dict([[str(h).lower(), r] for h, r in zip(header[3:], row[3:])])
if self.reference_id_equal_subject_id:
- kwargs["subject_id"] = reference_id
+ fields["subject_id"] = fields["reference_id"]
else:
- if "subject_id" not in kwargs:
+ if "subject_id" not in fields:
raise ValueError(f"`subject_id` not available in {header}")
- if "should_flip" not in kwargs:
+ if "should_flip" not in fields:
raise ValueError(f"`should_flip` not available in {header}")
+ if "purpose" not in fields:
+ raise ValueError(f"`purpose` not available in {header}")
+
+ kwargs = {k: fields[k] for k in fields.keys() - {"id",}}
+
+ # Retrieve the references list
+ if fields["purpose"].lower() == "enroll" and fields["reference_id"] not in self.references_list:
+ self.references_list.append(fields["reference_id"])
+ # Set the references list in the probes for vanilla-biometrics
+ if fields["purpose"].lower() != "enroll":
+ if fields["attack_type"]:
+ # Attacks only compare to the target (no `spoof_neg`)
+ kwargs["references"] = fields["reference_id"]
+ else:
+ kwargs["references"] = self.references_list
# One row leads to multiple samples (different frames)
all_samples = [DelayedSample(
functools.partial(
load_frame_from_file_replaymobile,
- file_name=os.path.join(self.dataset_original_directory, path + self.extension),
+ file_name=os.path.join(self.dataset_original_directory, fields["path"] + self.extension),
frame=frame,
should_flip=kwargs["should_flip"]=="TRUE",
),
- key=f"{id}_{frame}",
- path=path,
- reference_id=reference_id,
+ key=f"{fields['id']}_{frame}",
frame=frame,
**kwargs,
) for frame in range(12,251,24)]
return all_samples
+def read_frame_annotation_file_replaymobile(file_name, frame, annotations_type="json"):
+ """Returns the bounding-box for one frame of a video file of replay-mobile.
+
+ Given an annnotation file location and a frame number, returns the bounding
+ box coordinates corresponding to the frame.
+
+ The replay-mobile annotation files are composed of 4 columns and N rows for
+ N frames of the video:
+
+ 120 230 40 40
+ 125 230 40 40
+ ...
+ <x> <y> <w> <h>
+
+ Parameters
+ ----------
+
+ file_name: str
+ The annotation file name (relative to annotations_path).
+
+ frame: int
+ The video frame index.
+ """
+ logger.debug(f"Reading annotation file '{file_name}', frame {frame}.")
+
+ video_annotations = read_annotation_file(file_name, annotation_type=annotations_type)
+ # read_annotation_file returns an ordered dict with string keys
+ return video_annotations[f"{frame}"]
+
class FrameBoundingBoxAnnotationLoader(AnnotationsLoader):
"""A transformer that adds bounding-box to a sample from annotations files.
@@ -144,7 +155,7 @@ class FrameBoundingBoxAnnotationLoader(AnnotationsLoader):
"""
def __init__(self,
annotation_directory=None,
- annotation_extension=".face",
+ annotation_extension=".json",
**kwargs
):
super().__init__(
@@ -161,12 +172,7 @@ class FrameBoundingBoxAnnotationLoader(AnnotationsLoader):
annotated_samples = []
for x in X:
-
- # Build the path to the annotation files structure
- annotation_file = os.path.join(
- self.annotation_directory, x.path + self.annotation_extension
- )
-
+ # Adds the annotations as delayed_attributes, loading them when needed
annotated_samples.append(
DelayedSample(
x._load,
@@ -176,6 +182,7 @@ class FrameBoundingBoxAnnotationLoader(AnnotationsLoader):
read_frame_annotation_file_replaymobile,
file_name=f"{self.annotation_directory}:{x.path}{self.annotation_extension}",
frame=int(x.frame),
+ annotations_type=self.annotation_type,
)
),
)
@@ -183,6 +190,7 @@ class FrameBoundingBoxAnnotationLoader(AnnotationsLoader):
return annotated_samples
+
class ReplayMobileBioDatabase(CSVDataset):
"""Database interface that loads a csv definition for replay-mobile
@@ -258,7 +266,8 @@ class ReplayMobileBioDatabase(CSVDataset):
annotation_extension=annotations_extension,
),
),
+ fetch_probes=False,
**kwargs
)
- self.annotation_type = "bounding-box"
+ self.annotation_type = "eyes-center"
self.fixed_positions = None
--
GitLab